You are on page 1of 7

Consciousness and Complexity

Giulio Tononi and Gerald M. Edelman

Science 282, 1846 (1998);
DOI: 10.1126/science.282.5395.1846

If you wish to distribute this article to others, you can order high-quality copies for your
colleagues, clients, or customers by clicking here.
Permission to republish or repurpose articles or portions of articles can be obtained by
following the guidelines here.
The following resources related to this article are available online at (this information is current as of September 24, 2012 ):
Updated information and services, including high-resolution figures, can be found in the online
version of this article at:
This article cites 62 articles, 22 of which can be accessed free:
This article has been cited by 319 article(s) on the ISI Web of Science
This article has been cited by 42 articles hosted by HighWire Press; see:
This article appears in the following subject collections:

Science (print ISSN 0036-8075; online ISSN 1095-9203) is published weekly, except the last week in December, by the
American Association for the Advancement of Science, 1200 New York Avenue NW, Washington, DC 20005. Copyright
1998 by the American Association for the Advancement of Science; all rights reserved. The title Science is a
registered trademark of AAAS.

Downloaded from on September 24, 2012

This copy is for your personal, non-commercial use only.




Consciousness and Complexity

Conventional approaches to understanding consciousness are generally concerned
with the contribution of specific brain areas or groups of neurons. By contrast, it
is considered here what kinds of neural processes can account for key properties
of conscious experience. Applying measures of neural integration and complexity,
together with an analysis of extensive neurological data, leads to a testable
proposalthe dynamic core hypothesisabout the properties of the neural
substrate of consciousness.

hat is the neural substrate of conscious experience? While William

James concluded that it was the
entire brain (1), recent approaches have attempted to narrow the focus: are there neurons endowed with a special location or intrinsic property that are necessary and sufficient for conscious experience? Does primary
visual cortex contribute to conscious experience? Are brain areas that project directly to
prefrontal cortex more relevant than those
that do not (2)? Although heuristically useful,
these approaches leave a fundamental problem unresolved: How could the possession of
some particular anatomical location or biochemical feature render some neurons so
privileged that their activity gives rise to
subjective experience? Conferring this property on neurons seems to constitute a category error, in the sense of ascribing to things
properties they cannot have (3).
Here, we pursue a different approach. Instead of arguing whether a particular brain
area or group of neurons contributes to consciousness or not, our strategy is to characterize the kinds of neural processes that might
account for key properties of conscious experience. We emphasize two properties: conscious experience is integrated (each conscious scene is unified) and at the same time
it is highly differentiated (within a short time,
one can experience any of a huge number of
different conscious states). We first consider
neurobiological data indicating that neural
processes associated with conscious experience are highly integrated and highly differentiated. We then provide tools for measuring
integration (called functional clustering) and
differentiation (called neural complexity) that
are applicable to actual neural processes. This
leads us to formulate operational criteria for
The authors are at The Neurosciences Institute,
10640 John J. Hopkins Drive, San Diego, CA 92121,
*To whom correspondence should be addressed. Email:


determining whether the activity of a group

of neurons contributes to conscious experience. These criteria are incorporated into the
dynamic core hypothesis, a testable proposal
concerning the neural substrate of conscious
experience (4).

General Properties of Conscious

Consciousness, as William James pointed
out, is not a thing, but a process or stream that
is changing on a time scale of fractions of
seconds (1). As he emphasized, a fundamental aspect of the stream of consciousness is
that it is highly unified or integrated.
Integration. Integration is a property
shared by every conscious experience irrespective of its specific content: Each conscious state comprises a single scene that
cannot be decomposed into independent components (5). Integration is best appreciated by
considering the impossibility of conceiving
of a conscious scene that is not integrated,
that is, one which is not experienced from a
single point of view. A striking demonstration is given by split-brain patients performing a spatial memory task in which two independent sequences of visuospatial positions were presented, one to the left and one
to the right hemisphere (6 ). In these patients,
each hemisphere perceived a separate, simple
visual problem and the subjects were able to
solve the double task well. Normal subjects
could not treat the two independent visual
sequences as independent, parallel tasks. Instead, they combined the visual information
into a single conscious scene and into a single, large problem that was much more difficult to solve.
The unity of conscious experience is also
evidenced by our inability to perform multiple tasks, unless some tasks are highly automatic and impinge less on consciousness.
Moreover, we cannot make more than a single conscious decision within an interval of a
few hundreds of milliseconds, the so-called
psychological refractory period (7). Further-

more, we cannot be aware of two incongruent

scenes at the same time, as indicated by the
bistability of ambiguous figures and the phenomenon of perceptual rivalry (8). Unity also
entails that conscious experience is private,
that is, it is always experienced from a particular point of view and cannot fully be
shared (1).
Differentiation. While each conscious
state is an integrated whole, perhaps the most
remarkable property of conscious experience
is its extraordinary differentiation or complexity. The number of different conscious
states that can be accessed over a short time is
exceedingly large. For example, even if we
just consider visual images, we can easily
discriminate among innumerable scenes
within a fraction of a second (9). More generally, the occurrence of a given conscious
state implies an extremely rapid selection
among a repertoire of possible conscious
states that is, in fact, as large as ones experience and imagination. Differentiation
among a repertoire of possibilities constitutes
information, in the specific sense of reduction
of uncertainty (10). Although this is often
taken for granted, the occurrence of one particular conscious state over billions of others
therefore constitutes a correspondingly large
amount of information. Furthermore, it is information that makes a difference, in that it
may lead to different consequences in terms
of either thought or action.
The informativeness of consciousness
helps dispose of many of the paradoxes
raised about conscious experience. Consider a photodiode that can differentiate between light and dark and then provide an
audible output, and a conscious human performing the same task and giving a verbal
report. Why should the differentiation between light and dark performed by the human be associated with conscious experience, while presumably that performed by
the photodiode is not? The paradox disappears if one considers the information generated by such discriminations. To the photodiode, the discrimination between darkness and light is the only one available, and
is therefore minimally informative. To a
conscious human, by contrast, an experience of complete darkness and an experience of complete light are two specific
conscious experiences selected out of an
enormous repertoire, and their selection implies the availability of a correspondingly
large amount of information. To understand
consciousness, it is important to identify


Downloaded from on September 24, 2012

Giulio Tononi* and Gerald M. Edelman

underlying neural processes that are both

integrated and capable of such exceptionally informative differentiations.

General Properties of Neural

Processes Underlying Conscious
Distributed neural activity, particularly in the
thalamocortical system, is almost certainly
essential for determining the contents of conscious experience (4, 11). We suggested previously that a key neural mechanism underlying conscious experience are the reentrant
interactions between posterior thalamocortical areas involved in perceptual categorization and anterior areas related to memory,
value, and planning for action. Such interactions among neuronal groups in distributed
brain areas may be necessary in order to
generate a unified neural process corresponding to a multimodal conscious scene (4). Recent experimental findings are consistent
with this hypothesis and suggest some generalizations about the neural processes that
underlie conscious experience.
Activation and deactivation of distributed
neuronal populations. Changes in specific
aspects of conscious experience correlate
with changes in activity in specific brain
areas, whether the experience is driven by
external stimuli, by memory, or by imagery
and dreams (12). Conscious experience as
such, however, involves the activation or deactivation of widely distributed brain areas
(13), although what should count as the appropriate reference state for comparison is
not clear. In subjects who are comatose or
deeply anesthetized, unconsciousness is associated with a profound depression of neural
activity in both the cerebral cortex and thalamus (13). During slow-wave sleep, in which
consciousness is severely reduced or lost,
cerebral blood flow is globally reduced as
compared to both waking and REM (rapid
eye movement) sleep, two brain states associated with vivid conscious reports (14). A
more specific reference state would be the
response to a simple sensory input when a
subject is unaware of it versus when the
subject is aware of it. We have used magnetoencephalography to measure brain responses to flickering visual stimuli under conditions of binocular rivalry (15). A vertical
grating flickering at one frequency was presented to one eye and a horizontal grating,
flickering at a different frequency, was presented to the other eye. Although the stimuli
were presented together, the subjects perceived either the vertical grating or the horizontal grating, with an alternation every few
seconds. It was found that the power of
steady-state neuromagnetic responses at the
frequency of the flickering stimulus (its frequency tag) was higher by 30 to 60% in many
sensor locations when the subject was con-

scious of that stimulus. The sensors with

frequency tags that correlated with conscious
experience were widely distributed over both
posterior (occipital and temporal) and anterior (frontal) areas. Furthermore, there were
considerable variations among different subjects (Fig. 1).
A change in the degree to which neural
activity is distributed within the brain may
accompany the transition between conscious,
controlled performance and unconscious, automated performance. When tasks are novel,
brain activation related to the task is widely
distributed; when the task has become automatic, activation is more localized and may

Fig. 1. Amplitude and coherence differences

between the steady-state neuromagnetic responses during binocular rivalry when subjects
were conscious of a stimulus and when they
were not. The differences are taken between
amplitude and coherence values at 7.41 Hz
when the subjects were conscious of a vertical
grating flickered at 7.41 Hz and when they
were not (that is, when they were conscious of
a horizontal grating flickered at 9.5 Hz). Amplitude differences are topographically displayed
for two subjects. Color scale is in picotesla.
Significant positive differences in coherence at
7.41 Hz between pairs of distant sensors are
indicated by superimposed cyan lines. Blue
lines indicate negative differences in coherence.
Filled green circles indicate channels with signal-to-noise ratio .5 that have coherence values .0.3 with at least one other channel. See
(15) for details.

shift to a different set of areas (16). In animal

studies, neural activity related to sensory
stimuli can be recorded in many brain regions
before habituation. After habituation sets in
(a time when humans report that stimuli tend
to fade from consciousness), the same stimuli
evoke neural activity exclusively along their
specific sensory pathways (17). These observations suggest that when tasks are automatic
and require little or no conscious control, the
spread of signals that influence the performance of a task involves a more restricted
and dedicated set of circuits that become
functionally insulated. This produces a
gain in speed and precision, but a loss in
context-sensitivity, accessibility, and flexibility (18).
Integration through strong and rapid reentrant interactions. Activation and deactivation of distributed neural populations in the
thalamocortical system are not sufficient
bases for conscious experience unless the
activity of the neuronal groups involved is
integrated rapidly and effectively. We have
suggested that such rapid integration is
achieved through the process of reentrythe
ongoing, recursive, highly parallel signaling
within and among brain areas. Large-scale
computer simulations have shown that reentry can achieve the rapid integration or binding of distributed, functionally specialized
neuronal groups dynamically, that is, in a
unified neural process rather than in a single
place (19, 20).
Substantial evidence indicates that the integration of distributed neuronal populations
through reentrant interactions is required for
conscious experience. An indication comes
from the study of patients with disconnection
syndromes, in which one or more brain areas
are anatomically or functionally disconnected
from the rest of the brain due to some pathological process (21). In the paradigmatic disconnection syndrome (the split brain), visual
or somatosensory stimuli can activate the
nondominant hemisphere and lead to behavioral responses, but the dominant, verbal
hemisphere is not aware of them (22). Although the two hemispheres can still communicate through indirect, subcortical routes,
rapid and effective neural interactions mediated by direct reentrant connections are abolished by the lesion of the corpus callosum.
Modeling studies suggest that a telltale sign
of effective reentrant interactions is the occurrence of short-term temporal correlations
between the neuronal groups involved (19).
Experiments on cats show that short-term
temporal correlations between the activity of
neuronal groups responding to the same stimulus, but located in different hemispheres, are
abolished by callosal transections (23). Other
studies indicate that various kinds of cognitive tasks are accompanied by the occurrence
of short-term temporal correlations among SCIENCE VOL 282 4 DECEMBER 1998

Downloaded from on September 24, 2012



distributed populations of neurons in the

thalamocortical system (24). The magnetoencephalographic study of binocular rivalry
mentioned above (15) also indicates that
awareness of a stimulus is associated with
increased coherence among distant brain regions (Fig. 1).
The requirement for fast, strong, and distributed neural interactions may explain why
stimuli that are feeble, degraded, or shortlasting, often fail to be consciously perceived.
Although such stimuli may produce a behavioral response [perception without awareness
(25, 26 )], they are unlikely to ignite neural
activity of sufficient strength or duration to
support fast distributed interactions. Conversely, attention may increase the conscious
salience of certain stimuli by boosting the
corresponding neural responses as well as the
strength of neural interactions (27). Neural
activity is also more likely to contribute effectively to distributed neural interactions if it
is sustained for hundreds of milliseconds.
This would lead to the functional closure of
longer reentrant loops and thereby support
reentrant interactions among more distant regions (19, 20). Experimental findings are
consistent with this idea. High-frequency somatosensory stimuli delivered to the thalamus require about 500 ms for the production
of a conscious sensory experience, while less
than 150 ms are sufficient for sensory detection without awareness (28). The sustained
evoked potentials associated with a conscious
somatosensory sensation are apparently generated by the excitation of pyramidal neurons
of primary somatosensory cortex through reentrant interactions with higher cortical areas
Evidence for a correlation between conscious experience and sustained neural activity also comes from tasks involving visuospatial working memorythe ability to rehearse
or keep in mind a spatial location. Working
memory is used to bring or keep some item in
consciousness or close to conscious access
(30). In working memory tasks, sustained
neural activity is found in prefrontal cortex of
monkeys, and it is apparently maintained by
reentrant interactions between frontal and parietal regions (31). Sustained neural activity
may facilitate the integration of the activity of
spatially segregated brain regions into a coherent, multimodal neural process that is stable enough to permit decision-making and
planning (32).
Differentiated patterns of activity. Although strong and fast reentrant interactions
among distributed groups of neurons are necessary for conscious experience, in themselves, they are still not sufficient. This is
strikingly demonstrated by the unconsciousness accompanying generalized seizures and
slow-wave sleep. During generalized seizures, the brain is not only extremely active,


but most neurons fire in a highly synchronous

manner. For example, the electroencephalogram (EEG) during petit mal absences indicates that groups of neurons over the whole
brain are either all firing together or all silent
together, with these two neural states alternating every third of a second. Although such
hypersynchronous firing is indicative of
strong and distributed interactions, a subject
who is prey to such a seizure is unconscious.
Similarly, during slow-wave sleep, neurons
in the thalamocortical system are active as
well as remarkably interactive, as shown by
their synchronous firing in a stereotyped,
burst-pause pattern. During this stage of
sleep, however, it is rare to obtain vivid and
extensive conscious reports (33). By contrast,
during REM sleep, when neural activity is
not globally synchronous but resembles the
rapid and complex patterns of waking, subjects typically report vivid dreams if awakened. We suggest that the low-voltage, fastactivity EEG characteristic of waking and
REM sleep reflects the availability of a rich
and diverse repertoire of neural activity patterns. If the repertoire of differentiated neural
states is large, consciousness is possible.
Conversely, if this repertoire is reduced, as
when most groups of neurons in the cortex
discharge synchronously and functional discriminations among them are obliterated,
consciousness is curtailed or lost (34).

Theoretical Concepts and Measures

This brief review of neurological and neurophysiological data indicates that the distributed neural process underlying conscious experience must be functionally integrated and
at the same time highly differentiated. As
mentioned above, two key properties of conscious experience are that it is integrated, in
the sense that it cannot be subdivided into
independent components, and that it is extremely differentiated, in the sense that it is
possible, within a short time, to select among
an enormous number of different conscious
states. It is a central claim of this article that
analyzing the convergence between these
phenomenological and neural properties can
yield valuable insights into the kinds of neural processes that can account for the corresponding properties of conscious experience.
Such an analysis requires the availability of
satisfactory measures of integration and differentiation that can be applied to actual neural processes, as well as an understanding of
the neural mechanisms of integration.
Functional clustering: How to identify an
integrated process. How can one determine
whether a neural process is unified or simply
a collection of independent or nearly independent subprocesses? We have suggested
that a subset of distributed elements within a
system gives rise to a single, integrated process if, at a given time scale, these elements

interact much more strongly among themselves than with the rest of the system for
example, if they form a functional cluster.
This criterion has been formalized by introducing a direct measure of functional clustering (35) which we summarize here.
Consider a jth subset of k elements (Xkj)
taken from an isolated neural system X, and
its complement (X Xkj). Interactions between the subset and the rest of the system
introduce statistical dependence between the
two. This is measured most generally by their
mutual information MI(Xkj; X Xkj) 5 H(Xkj)
1 H(X Xkj) H(X), which captures the
extent to which the entropy of Xkj is accounted for by the entropy of X Xkj and vice versa
[H indicates statistical entropy (36)]. The
statistical dependence within a subset can be
measured by a generalization of mutual information, which is called integration and is
given by I(Xkj) 5 SH(xi) H(Xkj), where
H(xi) is the entropy of each element xi considered independently. We then define the
functional cluster index CI(Xkj) 5 I(Xkj)/
MI(Xkj; X Xkj) as a ratio of the statistical
dependence within the subset and the statistical dependence between that subset and the
rest of the system. Based on this definition, a
subset of neural elements that has a CI value
much higher than 1 and does not itself contain any smaller subset with a higher CI value
constitutes a functional cluster. This is a single, integrated neural process that cannot be
decomposed into independent or nearly independent components.
We have applied these measures of functional clustering both to simulated datasets
and to positron emission tomography data
obtained from schizophrenic subjects performing cognitive tasks (35). Theoretically
sound measures that can detect the occurrence of functional clustering at the time
scale (fractions of a second) crucial for conscious experience may require additional assumptions. Nevertheless, it would appear that
the rapid establishment of synchronous firing
among cortical regions and between cortex
and thalamus should be considered as an
indirect indicator of functional clustering,
since it implies strong and fast neural interactions among the neural populations involved (19, 20). The mechanisms of rapid
functional clustering among distributed populations of neurons in the thalamocortical
system have been studied with the help of
large-scale simulations (19, 20). These have
shown that the emergence of high-frequency
synchronous firing in the thalamocortical system depends critically on the dynamics of
corticothalamic and corticocortical reentrant
circuits and on the opening of voltage-dependent channels in the horizontal corticocortical
connections (37).
Neural complexity: Measuring the differences that make a difference. Once an inte-


Downloaded from on September 24, 2012


grated neural process is identified, we need to

determine to what degree that process is differentiated. Does it give rise to a large repertoire of different activity patterns or neural
states? It is essential to consider only those
differences between activity patterns that
make a difference to the system itself. A TV
screen may, for example, go through a large
number of activity patterns that look different to an external observer, but that make no
difference to the TV.
A possible approach to measuring differences that make a difference within an integrated neural system is to consider it as its
own observer. This can be achieved by
dividing the system (which, we assume, constitutes a functional cluster) into two subsets
and then measuring their mutual information
(38). The value of MI(Xkj; X Xkj) between
a jth subset Xkj of the isolated system X and
its complement X Xkj will be high if two
conditions are met. Both Xkj and X Xkj must
have many states [their entropy must be relatively high (10)], and the states of Xkj and of
X Xkj must be statistically dependent (the
entropy of Xkj must be largely accounted for
by the interactions with X Xkj, and vice
versa). The expression MI(Xkj; X Xkj) reflects how much, on average, changes in the
state of X Xkj make a difference to the state
of Xkj, and vice versa.
To obtain an overall measure of how differentiated a system is, one can consider not
just a single subset of its constituent elements, but all its possible subsets. The corresponding measure, called neural complexity,
is given by CN(X ) 5 S ^MI(Xkj; X Xkj)&,
where the sum is taken over all k subset sizes
and the average is taken over all jth combinations of k elements. Complexity is thus a
function of the average mutual information
between each subset and the rest of the system, and it reflects the number of states of a
system that result from interactions among its
elements (39).
It can be shown that high values of complexity reflect the coexistence of a high degree of functional specialization and functional integration within a system, as appears
to be the case for systems such as the brain.
For example, the dynamic behavior of a simulated cortical area containing thousands of
spontaneously active neuronal groups (38)
resembled the low-voltage fast-activity EEG
of waking states and had high complexity.
Such a system, whose connections were organized according to the rules found in the
cortex, visited a large repertoire of different
activity patterns that were the result of interactions among its elements. If the density of
the connections was reduced, the dynamic
behavior of the model resembled that of a
noisy TV screen and had minimal complexity. A large number of activity patterns were
visited, but they were merely the result of the

independent fluctuations of its elements. If

the connections within the cortical area were
instead distributed at random, the system
yielded a hypersynchronous EEG that resembled the high-voltage waves of slow-wave
sleep or of generalized epilepsy. The system
visited a very limited repertoire of activity
patterns, and its complexity was low.
Measures of complexity, like measures of
functional clustering, can also be applied to
neurophysiological data to evaluate the degree to which a neural process is both integrated and differentiated (40). This opens the
way to comparisons of the values of neural
complexity in different cognitive and arousal
states and to empirical tests of the relationships between brain complexity and conscious experience.

The Dynamic Core Hypothesis

A final issue we should consider is whether the
neural process underlying conscious experience
extends to most of the brain, as was concluded
by William James, or is restricted to varying
subsets of neuronal groups. Several observations support the latter possibility.
1) Classical lesion and stimulation studies
suggest that many brain structures outside the
thalamocortical system have no direct influence on conscious experience. Even within
the thalamocortical system, many regions can
be lesioned or stimulated without producing
direct effects on conscious experience (41).
2) Neurophysiological studies indicate a
possible dissociation between conscious experience and ongoing neural activity within
portions of the thalamocortical system. During binocular rivalry in monkeys, a large
proportion of neurons in early visual areas,
such as V1, V4, and MT, continued to fire to
their preferred stimulus even when it was not
consciously perceived (42). The activity of
only a subset of the neurons recorded in these
areas was correlated with the percept, although in higher areas such as IT and STS,
the percentage reached 95%. In our magnetoencephalographic study of binocular rivalry
in humans (Fig. 1) (15), we found that the
responses of only a subset of occipital, temporal, and frontal areas was correlated with
the conscious perception of a stimulus, although several other regions showed widespread responses to stimuli that were not
consciously perceived.
3) The firing of neurons dealing with rapidly varying local details of a sensory input or
a motor output does not seem to map to
conscious experience. The latter deals with
invariant properties of objects that are highly
informative as well as more stable and easily
manipulated. For example, patterns of neural
activity in the retina and other early visual
structures correspond faithfully to spatial and
temporal details of the visual input and are in
constant flux. During each visual fixation,

however, humans extract the meaning of a

scene and are not conscious of considerable
changes in its local details (43). Groups of
neurons responding in a stable way to invariant properties of objects are therefore more
likely to contribute to conscious experience.
4) Many neural processes devoted to carrying out highly automated routines that
make it possible to talk, listen, read, write,
and so forth, in a fast and effortless way do
not appear to contribute directly to conscious
experience, although they are essential in determining its content (44). As mentioned
above, neural circuits carrying out such highly practiced neural routines may become
functionally insulated except at the input or
output stages. There is also some evidence
that cortical regions that are part of a fast
system for controlling action, such as the
dorsal visual stream, may not contribute significantly to conscious experience (45).
5) Although the sheer anatomical connectivity of the brain may hint that, over a sufficiently long time scale, everything can interact with everything else, modeling studies
indicate that only certain interactions within
the thalamocortical system are fast and strong
enough to lead to the formation of a large
functional cluster within a few hundred milliseconds (46).
These observations suggest that changes
in the firing of only certain distributed subsets of the neuronal groups that are activated
or deactivated in response to a given task are
associated with conscious experience. What
is special about these subsets of neuronal
groups, and how can they be identified? We
suggest the following:
1) A group of neurons can contribute
directly to conscious experience only if it is
part of a distributed functional cluster that
achieves high integration in hundreds of
2) To sustain conscious experience, it is
essential that this functional cluster be highly
differentiated, as indicated by high values of
We propose that a large cluster of neuronal groups that together constitute, on a time
scale of hundreds of milliseconds, a unified
neural process of high complexity be termed
the dynamic core, in order to emphasize
both its integration and its constantly changing activity patterns. The dynamic core is a
functional cluster: its participating neuronal
groups are much more strongly interactive
among themselves than with the rest of the
brain. The dynamic core must also have high
complexity: its global activity patterns must
be selected within less than a second out of a
very large repertoire.
The dynamic core would typically include
posterior corticothalamic regions involved in
perceptual categorization interacting reentrantly
with anterior regions involved in concept for- SCIENCE VOL 282 4 DECEMBER 1998

Downloaded from on September 24, 2012





changes in the functional connectivity among

distributed populations of neurons can occur
independently of firing rate (50). Recent
studies in monkey frontal cortex also show
abrupt and simultaneous shifts among stationary activity states involving several, but
not all recorded neurons (51). A convincing
demonstration of rapid functional clustering
among distributed neuronal groups requires,
however, that these studies be extended to
larger populations of neurons in several brain
areas. Another possibility would be to examine whether the effects of direct cortical microstimulation spread more widely in the
brain if they are associated with conscious
experience than if they are not. In humans,
the extent and boundaries of neural populations exchanging coherent signals can be
evaluated through methods of frequency tagging (15). Techniques offering both wide
spatial coverage and high temporal resolution
could also help establish how large a dynamic
core normally is, how its composition changes, and whether certain brain regions are always included or always excluded. It is also
significant to ask whether the dynamic core
can split, and thus whether multiple dynamic
cores can coexist in a normal subject. A
reasonable prediction would be that certain
disorders of consciousness, notably dissociative disorders and schizophrenia, should be
reflected in abnormalities of the dynamic
core and possibly result in the formation of
multiple cores.
A strong prediction based on our hypothesis is that the complexity of the dynamic
core should correlate with the conscious state
of the subject. For example, we predict that
neural complexity should be much higher
during waking and REM sleep than during
the deep stages of slow-wave sleep, and that
it should be extremely low during epileptic
seizures despite the overall increase in brain
activity. We also predict that neural processes
underlying automatic behaviors, no matter
how sophisticated, should have lower complexity than neural processes underlying consciously controlled behaviors. Finally, a systematic increase in the complexity of coherent neural processes is expected to accompany cognitive development.
The outcome of such tests should indicate
whether conscious phenomenology can indeed be related, as we suggest, to a distributed neural process that is both highly integrated and highly differentiated. The evidence available so far supports the belief that
a scientific explanation of consciousness is
becoming increasingly feasible (52).
References and Notes

1. W. James, The Principles of Psychology (Holt, New

York, 1890).
2. F. Crick and C. Koch, Cold Spring Harbor Symp.
Quant. Biol. 55, 953 (1990); Nature 375, 121 (1995);










S. Zeki and A. Bartels, Proc. R. Soc. London Ser. B 265,

1583 (1998).
G. Ryle, The Concept of Mind (Hutchinson, London,
G. M. Edelman, The Remembered Present (Basic
Books, New York, 1989);
and G. Tononi,
Consciousness: How Matter Becomes Imagination
(Basic Books, New York, in press); see also G. Tononi
and G. M. Edelman, in Consciousness, H. Jasper et al.,
Eds. (Plenum, New York, 1998). pp. 245280.
A conscious state is meant here as an idealization,
exemplified by viewing a rapid succession of slides.
J. D. Holtzman and M. S. Gazzaniga, Neuropsychologia 23, 315 (1985).
H. Pashler, Psychol. Bull. 116, 220 (1994). The duration of this interval is comparable with the duration
of conscious states [A. L. Blumenthal, The Process
of Cognition (Prentice-Hall, Englewood Cliffs, NJ,
F. Sengpiel, Curr. Biol. 7, R447 (1997).
H. Intraub, J. Exp. Psychol. Hum. Percept. Perform. 7,
604 (1981); I. Biederman, Science 177, 77 (1972).
C. E. Shannon and W. Weaver, The Mathematical
Theory of Communication (Univ. of Illinois Press, Urbana, IL, 1963). Note that the informativeness of
consciousness also helps us to understand its evolutionary value (4).
V. B. Mountcastle, in The Mindful Brain, G. M. Edelman and V. B. Mountcastle, Eds. (MIT Press, Cambridge, MA, 1978), p. 7; A. Damasio, Cognition 33, 25
(1989); R. Llinas, U. Ribary, M. Joliot, X.-J. Wang, in
Temporal Coding in the Brain, G. Buzsaki, R. Llinas, W.
Singer, Eds. (Springer-Verlag, Berlin, 1994); J. Newman, Consciousness Cognit. 4, 172 (1995); T. W.
Picton and D. T. Stuss, Curr. Biol. 4, 256 (1994).
R. S. J. Frackowiak, Human Brain Function (Academic Press, San Diego, CA, 1997); P. E. Roland, Brain
Activation ( Wiley-Liss, New York, 1993); M. I. Posner and M. E. Raichle, Images of Mind (Scientific
American Library, New York, 1994). These imaging
studies confirm and extend previous lesion and
stimulation studies.
Lesion studies indicate that consciousness is abolished by widely distributed damage but not by localized cortical damage. The only localized brain lesions
resulting in loss of consciousness typically affect the
reticular core in the upper brainstem and hypothalamus or its rostral extensions in the reticular and
intralaminar thalamic nuclei [F. Plum, in Normal and
Altered States of Function, A. Peters and E. G. Jones,
Eds. (Plenum, New York, 1991), vol. 9, p. 359]. Although it has been suggested that the reticular core
may have a privileged connection to conscious experience [ J. E. Bogen, Consciousness Cognit. 4, 52
(1995)], its activity may simply be required to sustain
distributed activity patterns in the cortex.
A. R. Braun et al., Science 279, 91 (1998); P. Maquet
et al., Nature 383, 163 (1996). Neural activity in
slow-wave sleep is reduced in both anterior neocortical regions (most of the prefrontal cortex), as well
as in posterior cortical regions (especially parietal
association areas), in paralimbic structures (anterior
cingulate cortex and anterior insula), and in centrencephalic structures (reticular activating system, thalamus, and basal ganglia); in contrast, it is not depressed in unimodal sensory areas (primary visual,
auditory, and somatosensory cortex).
G. Tononi, R. Srinivasan, D. P. Russell, G. M. Edelman,
Proc. Natl. Acad. Sci. U.S.A. 95, 3198 (1998); R.
Srinivasan, D. P. Russell, G. M. Edelman, G. Tononi,
Soc. Neurosci. Abstr. 24, 433 (1998).
S. E. Petersen, H. vanMier, J. A. Fiez, M. E. Raichle,
Proc. Natl. Acad. Sci. U.S.A. 95, 853 (1998); R. J.
Haier et al., Brain Res. 570, 134 (1992).
J. A. Horel et al., Science 158, 394 (1967).
B. J. Baars, A Cognitive Theory of Consciousness (Cambridge Univ. Press, New York, 1988).
O. Sporns, G. Tononi, G. M. Edelman, Proc. Natl.
Acad. Sci. U.S.A. 88, 129 (1991); G. Tononi, O.
Sporns, G. M. Edelman, Cereb. Cortex 2, 310 (1992).
E. D. Lumer, G. M. Edelman, G. Tononi, Cereb. Cortex
7, 207 (1997); ibid., p. 228. For example, in a largescale model of the visual system, reentrant interactions between groups of neurons in perceptual or
posterior areas and in executive or anterior areas



Downloaded from on September 24, 2012

mation, value-related memory, and planning

(4), although it would not necessarily be restricted to the thalamocortical system. The term
dynamic core deliberately does not refer to a
unique, invariant set of brain areas (be they
prefrontal, extrastriate, or striate cortex), and
the core may change in composition over time
(47). Because our hypothesis highlights the role
of the functional interactions among distributed
groups of neurons rather than their local properties (2), the same group of neurons may at
times be part of the dynamic core and underlie
conscious experience, while at other times it
may not be part of it and thus be involved in
unconscious processes. Furthermore, since participation in the dynamic core depends on the
rapidly shifting functional connectivity among
groups of neurons rather than on anatomical
proximity, the composition of the core can transcend traditional anatomical boundaries (48).
Finally, as suggested by imaging studies (15),
the exact composition of the core related to
particular conscious states is expected to vary
significantly across individuals.
The dynamic core hypothesis avoids the
category error of assuming that certain local,
intrinsic properties of neurons have, in some
mysterious way, a privileged correlation with
consciousness. Instead, this hypothesis accounts for fundamental properties of conscious experience by linking them to global
properties of particular neural processes. We
have seen that conscious experience is a process that is unified and private, that is extremely differentiated, and that evolves on a
time scale of hundreds of milliseconds. The
dynamic core is a process, since it is characterized in terms of time-varying neural interactions, not as a thing or a location. It is
unified and private, because its integration
must be high at the same time as its mutual
information with what surrounds is low, thus
creating a functional boundary between what
is part of it and what is not. The requirement
for high complexity means that the dynamic
core must be highly differentiatedit must
be able to select, based on its intrinsic interactions, among a large repertoire of different
activity patterns. Finally, the selection among
integrated states must be achieved within
hundreds of milliseconds, thus reflecting the
time course of conscious experience (49).
A number of experimental questions and
associated predictions are generated by this
hypothesis. A central prediction is that, during cognitive activities involving consciousness, there should be evidence for a large but
distinct set of distributed neuronal groups that
interact over fractions of a second much more
strongly among themselves than with the rest
of the brain. This prediction could, in principle, be tested by recording, in parallel, multiple neurons whose activity is correlated
with conscious experience. Multielectrode recordings have already indicated that rapid







rapidly led to their synchronous firing and to a correct behavioral discrimination. This discrimination
was based on the dynamic binding of multiple visual
attributes (position, movement, color, form) and of
different levels of stimulus generalization (local features, invariant aspects of stimuli).
B. Kolb and I. Q. Whishaw, Fundamentals of Human
Neuropsychology (Freeman, New York, 1996). Psychiatric dissociation syndromes and conversion disorders may originate from a similar alteration of
reentrant interactions, although in these cases, the
disconnection would be functional rather than anatomical [ J. F. Kihlstrom, Consciousness Cognit. 1, 47
(1992)]. Some explicit-implicit dissociations, such as
amnesia, may also be due to a partial disconnection
of a lesioned area from the more global pattern of
neural activity that is associated with consciousness
[D. L. Schacter, Proc. Natl. Acad. Sci. U.S.A. 89, 11113
M. S. Gazzaniga, Neuron 14, 217 (1995).
A. K. Engel, P. Konig, A. K. Kreiter, W. Singer, Science
252, 1177 (1991).
S. L. Bressler, Brain Res. Rev. 20, 288 (1995); W.
Singer and C. M. Gray, Annu. Rev. Neurosci. 18, 555
(1995); M. Joliot, U. Ribary, R. Llinas, Proc. Natl. Acad.
Sci. U.S.A. 91, 11748 (1994); A. Gevins et al., Electroencephalogr. Clin. Neurophysiol. 98, 327 (1996).
A. J. Marcel, Cognit. Psychol. 15, 238 (1983); ibid., p.
197; P. M. Merikle, Am. Psychol. 47, 792 (1992). In
some cases, perception without awareness has been
shown to occur with stimuli that are not short-lasting
or weak [F. C. Kolb and J. Braun, Nature 377, 336
(1995); S. He, H. S. Smallman, D. I. A. MacLeod, Invest.
Ophthalmol. Visual Sci. 36, S438 (1995).
S. He, P. Cavanagh, J. Intriligator, Nature 383, 334
J. H. Maunsell, Science 270, 764 (1995); K. J. Friston,
Proc. Natl. Acad. Sci. U.S.A. 95, 796 (1998).
B. Libet, Ciba Found. Symp. 174, 123 (1993).
L. Cauller, Behav. Brain Res. 71, 163 (1995).
A. Baddeley, Proc. Natl. Acad. Sci. U.S.A. 93, 13468
J. M. Fuster, R. H. Bauer, J. P. Jervey, Brain Res. 330,
299 (1985); P. S. Goldman-Rakic and M. Chafee, Soc.
Neurosci. Abstr. 20, 808 (1994).
The idea that neural activity must persist for a minimum period of time in order to contribute to conscious experience is also suggested by the phenomenon of masking [ J. L. Taylor and D. I. McCloskey, Exp.
Brain Res. 110, 62 (1996); K. J. Meador et al., Neurology 51, 721 (1998)].
M. Steriade, Cereb. Cortex 7, 583 (1997); D. Kahn,
E. F. Pace-Schott, J. A. Hobson, Neuroscience 78, 13
Neural activity must also exhibit sufficient variance
in time to support conscious perception. For example,
if images on the retina are stabilized, perception
fades rapidly, and a similar effect is seen in Ganzfeld
stimulation. Short-lasting visual stimuli become invisible if the transient neuronal responses associated
with their onset and offset are suppressed by masking stimuli [S. L. Macknik and M. S. Livingstone,
Nature Neurosci. 1, 144 (1998)].
G. Tononi, A. R. McIntosh, D. P. Russell, G. M. Edelman, Neuroimage 7, 133 (1998).
As a measure of statistical dependence, mutual information has the virtue of being highly general,
because it is multivariate and sensitive to high-order
moments of statistical dependence [A. Papoulis,
Probability, Random Variables, and Stochastic Processes (McGraw-Hill, New York, 1991)]. Note that
mutual information reflects a statistical dependence
among subsets of a system, irrespective of its source.
The presence and direction of causal interactions
between two subsets of a system can be evaluated,








at least in principle, by measuring the change in

mutual information obtained by perturbing or deefferenting each subset in turn.
These observations are of interest in view of the
well-known action of certain so-called dissociative
anesthetics, such as ketamine and phencyclidine, that
act as noncompetitive antagonists of the N-methylD-aspartate receptor [H. Flohr, Behav. Brain Res. 71,
157 (1995)].
G. Tononi, O. Sporns, G. M. Edelman, Proc. Natl.
Acad. Sci. U.S.A. 91, 5033 (1994). A complexity
measure that does not involve the calculation of
average values of integration and mutual information
can also be defined as the amount of the entropy of
a system that is accounted for by the interactions
among its elements and is given by SMI(Xj1; X Xj1)
I(X) [G. Tononi, G. M. Edelman, O. Sporns, Trends
Cognit. Sci., in press]. Note that complexity measures
should be applied to a single system (a functional
cluster) and not to a collection of independent or
nearly independent subsystems.
Changes in complexity can be obtained without modifying the anatomical connectivity of the model by
simulating the transition between the burst-pause
pattern of firing typical of slow-wave sleep and the
tonic mode of firing typical of waking and REM sleep
(G. Tononi, unpublished material). It should be noted
that high complexity is not easy to achieve. A system
of elements that are randomly interconnected, for
instance, may look very complicated, but it has low
values of complexity. On the other hand, systems
that undergo selective processes so as to match the
statistical structure of a rich environment will gradually increase their complexity [G. Tononi, O. Sporns,
G. M. Edelman, Proc. Natl. Acad. Sci. U.S.A. 93, 3422
K. J. Friston, G. Tononi, O. Sporns, G. M. Edelman,
Hum. Brain Mapp. 3, 302 (1995).
W. Penfield, The Excitable Cortex in Conscious Man
(Thomas, Springfield, IL, 1958).
D. A. Leopold and N. K. Logothetis, Nature 379, 549
(1996); D. L. Shenberg and N. K. Logothetis, Proc.
Natl. Acad. Sci. U.S.A. 94, 3408 (1997). For other
instances of dissociation, see (26); M. Gur and D. M.
Snodderly, Vision Res. 37, 377 (1997); I. N. Pigarev,
H. C. Nothdurft, S. Kastner, Neuroreport 8, 2557
(1997); D. C. Bradley, G. C. Chang, R. A. Andersen,
Nature 392, 714 (1998).
D. J. Simons and D. T. Levin, Trends Cogn. Sci. 1, 261
(1997). The neurological evidence is in agreement
with these psychological observations. In the adult,
lesions of the retina produce blindness, but they do
not eliminate visual imagery, visual memories, and
visual dreams, while the latter are eliminated by
lesions of certain visual cortical areas [M. Solms, The
Neuropsychology of Dreams (Erlbaum, Mahwah, NJ,
1997)]. V1 may be important, however, to provide
visual consciousness with a certain degree of detail.
See also R. Jackendoff [Consciousness and the Computational Mind (MIT Press, Cambridge, MA, 1987)].
R. M. Shiffrin, in Scientific Approaches to Consciousness, J. D. Cohen and J. W. Schooler, Eds. (Erlbaum,
Mahwah, NJ, 1997), p. 49; L. L. Jacoby, D. Ste-Marie,
J. P. Toth, in Attention: Selection, Awareness, and
Control, A. D. Baddeley and L. Weiskrantz, Eds. (Clarendon, Oxford, 1993), p. 261; W. Schneider, M.
Pimm-Smith, M. Worden, Curr. Opin. Neurobiol. 4,
177 (1994).
A. D. Milner, Neuropsychologia 33, 1117 (1995);
and M. A. Goodale, The Visual Brain in Action
(Oxford Univ. Press, New York, 1995).
The organization of the anatomical connectivity of
certain brain regions, such as the thalamocortical
system, is much more effective in generating coherent dynamic states than that of other regions, such as






the cerebellum or the basal ganglia (G. Tononi, unpublished material). Consistent with this, although in
cortical and thalamic areas 20 to 50% of all pairs of
neurons recorded are broadly synchronized, neurons
in the internal segment of the globus pallidus, the
output station of the basal ganglia, are almost completely uncorrelated [H. Bergman et al., Trends Neurosci. 21, 32 (1998)].
If the fast integration of neural activity comes at a
premium in terms of number of connections and
energetic requirements, neuronal groups in higher
areas should be privileged members of the dynamic
core underlying consciousness. Everything else being
equal, their firing is more informative, in the sense
that it rules out a larger number of possibilities. For
example, the firing of face-selective neurons in area
IT considerably reduces uncertainty about a visual
scene (seeing a face rules out countless other visual
scenes), while the firing of retinal neurons reduces
uncertainty by much less (a bright spot in a certain
position of the visual field is consistent with countless visual scenes). The results of studies of binocular
rivalry in monkeys and humans mentioned above are
consistent with this view.
We emphasize that the dynamic core, the highly
complex, rapidly established functional cluster proposed to underlie conscious experience, is in no way
the only integrated but distributed neural process
that is relevant to brain function. We have hypothesized that distributed but integrated neural processes called global mappings, encompassing portions of
the thalamocortical system, as well as parallel loops
through cortical appendages such as the basal ganglia, the hippocampus, and the cerebellum, underlie
the unity of behavioral sequences (4). The functional
integration of global mappings is envisioned to occur
at longer time scales than the dynamic core (seconds
as opposed to fractions of a second). However, these
two kinds of dynamic processes are expected to
partially overlap for short periods of time.
Qualiathe seemingly inexplicable phenomenological manifestations of conscious experienceare conceived within this framework as rapid, highly informative discriminations within a repertoire of billions
of neural states available to a unified neural process
of great complexity. They correspond to the generation of a large amount of information in a short
period of time. In this view, each quale even a
seemingly simple quale like a feeling of redness
corresponds to a discriminable state of the dynamic
core in its entirety, and not merely to the state of a
specific group of neurons in a certain brain area. The
subjective meaning or quale of redness, for example, would be defined by the (increased) activity of
red-selective neurons as much as by the (reduced or
unmodified) activity of neuronal groups selective for
green or blue, for visual motion or shape, for auditory
or somatosensory events, and for proprioceptive inputs, body schemas, emotions, intentions, and so
forth, that jointly constitute the dynamic core. This
view is antithetical to modular or atomistic approaches to consciousness (2).
E. Vaadia et al., Nature 373, 515 (1995).
E. Seidemann, I. Meilijson, M. Abeles, H. Bergman, E.
Vaadia, J. Neurosci. 16, 752 (1996).
It is perhaps worth pointing out that our analysis
predicts the possibility of constructing a conscious
artifact and outlines some key principles that should
constrain its construction. This work was carried out
as part of the theoretical neurobiology program at
The Neurosciences Institute, which is supported by
Neurosciences Research Foundation. The Foundation
receives major support for this program from Novartis Pharmaceuticals Corporation and the W. M. Keck
Foundation. SCIENCE VOL 282 4 DECEMBER 1998

Downloaded from on September 24, 2012